Speech emotion recognition via graph-based representations.

Area of Science:

• Affective computing
• Speech processing
• Machine learning

Background:

• Speech emotion recognition (SER) is crucial for affective computing.
• Existing SER methods utilize diverse features, algorithms, and datasets.
• A need exists for novel feature extraction techniques in SER.

Purpose of the Study:

• To apply graph theory for classifying emotionally-colored speech signals.
• To develop a novel feature set using statistical and structural information from time series data.
• To establish unique feature-based identities for emotions within each speaker.

Main Methods:

• Graph theory was employed to extract novel features from speech signals.
• A Random Forest classifier was used for emotion classification.
• Leave-One-Speaker-Out Cross Validation (LOSO-CV) was implemented for robust evaluation.

Main Results:

• The proposed graph theory-based method outperformed two state-of-the-art approaches.
• Superior performance was observed on EMODB, AESDD, and DEMoS datasets.
• Significant average Unweighted Average Recall (UAR) increases were noted across datasets.

Conclusions:

• Graph theory offers a powerful tool for novel feature extraction in SER.
• The proposed method demonstrates superior performance compared to traditional and deep learning approaches.
• This research advances SER capabilities by providing a more effective feature representation.